The Log is the Agent: An Event-Sourcing-Based Reactive Graph for Auditable and Forkable Agent Systems

Existing AI agent frameworks typically build an LLM-centric conversation loop first, then add a logging layer afterward for observability. In contrast, ActiveGraph, proposed in this paper, inverts this structure by making the append-only event log the sole Source of Truth for the system.

The key features of this architecture are as follows:

• Deterministic Projection: The operating graph is a deterministic artifact generated from the event log.
• Reactive Collaboration: Without direct commands between components, functions, classes, and LLM routines react to changes in the graph and generate new events.
• Three Core Benefits:
  • Deterministic Replay: The entire execution process can be perfectly reproduced through the log.
  • Cheap Forking: Branches can be created at intermediate execution points to test various scenarios without re-execution.
  • End-to-end Lineage: All causal relationships can be traced, from high-level goals down to individual model calls.

This design provides a highly advantageous foundation for building self-improving mechanisms in agents.